Beijing-Moore Threads 2022 Autumn Conference was held successfully on November 3, 2022 in Beijing Zhongguancun National Independent Innovation Demonstration Zone. Moore Threads announced a number of innovative achievements at the press conference, including a new multi-functional GPU product and MUSA software stack, demonstrating the comprehensive strength of Moore Thread’s “combination of software and hardware.” MTT S3000, a server-oriented multifunctional GPU product based on the MUSA architecture, is one of them. It supports popular graphics and computing interfaces like DirectX, OpenGL, OpenGL ES, Vulkan, and OpenCL. It is CUDA-compatible and can be used for AI reasoning and training, cloud games, and cloud rendering, as well as other scenarios to provide general intelligent computing power support, with the goal of laying a solid computing power foundation for the construction of data centers, intelligent computing centers, and meta-computing centers, as well as facilitating the innovation and implementation of metaverse multiple applications.
Moore Thread’s new second multi-functional GPU chip “Chunxiao” powers the MTT S3000, which includes 4096 MUSA stream processing cores and 128 dedicated tensor computing cores. The number of transistors reaches 22 billion, the operating frequency is 1.9GHz, and the memory is 256bit wide; with 32GB GDDR6 memory; supports FP32, FP16, INT8, and other computing precision, with FP32 computing power reaching 15.2TFLOPS.
MTT S3000 is the first product in the GPU industry to support the PCIe 5.0 standard, as well as the first to adopt the PCIe 5.0 interface. The increased bandwidth provided by PCIe 5.0 can improve not only communication efficiency between GPU and CPU, but also inter-core communication bandwidth during “multi-card” deployment, improving overall computing efficiency and performance of GPU clusters.
MTT S3000 has an integrated MUSA intelligent multimedia engine 2.0 and hardware virtualization function, which can expand the graphics card’s application range and the user’s application experience in multiple dimensions. Simultaneously, the MTT S3000 offers two sets of DP 1.4a display output interfaces. The increased display output capability will increase the MTT S3000’s business adaptability.
The evolution of data centers to intelligent computing centers and meta-computing centers is the dominant trend in the industry’s development of computing power demand. The high demand for GPU computing power from AI and metaverse application loads is the primary force driving this evolution trend. With the support of the MUSA software and hardware computing platform, the Moore Threads MTT S3000, which is specially designed to improve intelligent computing power, can achieve an average improvement of four times in AI application performance compared to “Sudi.”
MTT S3000 and its supporting software and hardware products cover the entire process, from algorithm model to application deployment, and can offer AI users a user-friendly and comprehensive package solution. Moore Threads can provide digital humans, digital customer service, content generation, and other industry solutions at the algorithm level, as well as a rich pre-training model library to empower AI developers and improve development efficiency; at the application solution level, Moore Threads can provide digital humans, digital customer service, content generation, and other industry solutions, serving financial insurance, education, medical, and other related industry users.
In terms of deep learning training, the MTT S3000 has a number of advantages, including ease of use, scalability, and compatibility. MTT S3000, which is based on the MUSA software stack, can fully support existing algorithms and can support multiple training modes such as single-machine single-card, single-machine multi-card, and multi-machine multi-card. Not only that, but MTT S3000 is also compatible with a variety of popular deep learning frameworks, including PyTorch, TensorFlow, Baidu PaddlePaddle, and Jittor, and can optimize dozens of AI models such as Transformer, CNN, and RNN.
MTT S3000 supports mainstream AI models in vision, speech, natural speech understanding, and multimodality when it comes to deep learning reasoning. Moore Threads also continued to optimize the MUSA software stack’s in-depth performance and launched TensorX, a self-developed AI inference engine that greatly improved inference performance by utilizing technologies such as task concurrency, adaptive algorithm optimization, memory access optimization, and operator optimization. The MTT S3000 can meet the needs of industries that emphasize high-precision reasoning, such as biomedicine, finance, and insurance, while also providing extreme performance.
Simultaneously, with the help of Moore Threads’ CUDA ON MUSA compatibility solution, users can seamlessly migrate CUDA-developed code to MTT S3000.
GPU is the primary computing power required by today’s data centers, and cloud-native technology applications in cloud computing, cloud desktops, cloud games, and other fields are becoming more common. Moore Threads has released a series of GPU cloud-native solutions based on the innovative MT Mesh 2.0.
MT Mesh 2.0 can automatically allocate GPU computing and memory resources based on the cloud center’s application load, allowing for elastic scaling of GPU computing power. It can not only split a GPU across multiple containers or virtual machines at will, but it can also schedule multiple GPUs within a container or virtual machine.
Elastic containerized GPU (sGPU): Use MT Mesh 2.0 to achieve flexible and effective deployment of containerized GPU resources based on the Kubernetes ecosystem.
Elastic virtualized GPU (vGPU): It is the first in the industry to create resource elastic segmentation technology using MT Mesh 2.0. It is capable of dynamically allocating and modifying GPU virtualized resources without restarting, as well as achieving computing power on demand, dynamic scaling, and exhaustion. Release; add a new “time-space segmentation” feature, support hardware virtualization (SR-IOV), secure physical segmentation, up to 32-way virtualization, Windows cloud desktop GPU virtualization, and GPU pass-through in Tongxin and Kirin operating systems.
Cloud acceleration technology for Android containers ACX: Using GPU Android container transparent transmission technology and rendering and encoding integration technology, Android cloud mobile phone solutions can be accelerated, application delays reduced, and the number of concurrent channels increased. It can provide better Android cloud game rendering effects and compatibility features by natively supporting OpenGL ES, OpenGL rendering framework, and hardware acceleration of Android game material compression algorithms such as ETC and ETC2.
MTT S3000 has powerful video processing capabilities and integrates Moore’s thread second-generation intelligent multimedia engine (MT Smart Media Engine 2.0), which can accelerate various audio and video applications in the cloud, such as video AI analysis, video conferencing, cloud games, and cloud desktops and real-time interaction, and can support up to 8K ultra-high-definition HDR video real-time processing, and can support up to 8K ultra-high-definition HDR video real-time processing.
MTT S3000 supports popular audio and video frameworks such as FFmpeg and GStreamer, as well as general video acceleration APIs across a variety of operating systems. Moore Threads in the multi-functional GPU combines the intelligent multimedia engine, graphics engine, and AI computing engine to provide a rich SDK development kit that greatly reduces CPU load and data copying and is easy for developers to call. High-definition transcoding, intelligent content generation, AI image quality enhancement, old film restoration, digital twin, digital human, and other solutions offer dependable assistance.
Moore Threads also unveiled the MTVERSE metaverse platform, as well as a slew of software and hardware products, including the MCCX metacomputing all-in-one computer based on the MTT S3000, at the press conference.
MTVERSE uses the Moore threads MUSA GPU cluster as the computing power foundation to provide users with computing infrastructure and services, including three big data platforms, AI training and reasoning, graphics rendering, and physical simulation, as well as hardware clusters, software infrastructure, and SDK tool chains. Its full-stack solution addresses multiple core elements in the Metaverse, including people, scenes, and content. The massive SDK tools on the upper layer can assist developers and applications in easily invoking these capabilities to realize a variety of functions such as digital human, document understanding, speech recognition, visual recognition, natural semantic understanding, dialogue interaction, physical simulation, AIGC content generation, and so on, further simplifying application and solution development and shortening the development cycle.
Metaverse will guide the future development of the global digital economy. It is still in its early stages, but with the entry of many business titans and successive policy support from the governments of various parts of China, the prospects are infinitely broad. In the future, the metaverse must be an integrated application with graphics and AI computing as the foundation. The development of physics engines, rendering engines, and AI technology heavily influences the rate at which metaverse applications evolve. Moore Threads will continue to maintain a high-speed iterative innovation model of GPU products, improve software and hardware capabilities, and promote metaverse application innovation through extensive ecological cooperation, ultimately serving the high-quality development of the digital economy.
Moore Threads is a high-tech integrated circuit company that specializes in GPU chip design. Its primary focus is the development and design of full-featured GPU chips and related products that can provide powerful computing acceleration capabilities to Chinese technology partners. Founded in October 2020, the company is dedicated to developing a new generation of GPUs for meta-computing applications, constructing a comprehensive computing platform that integrates visual computing, 3D graphics computing, scientific computing, and artificial intelligence computing, and establishing an ecosystem based on cloud-native GPU computing to help drive the digital economy’s development.
Moore Threads is distributed by Unixcloud Technology (Shenzhen) Co., Ltd. Unixcloud has a long history of product development and manufacturing. Moore Threads’ official authorized distributor is UnixCloud. It focuses on the field of edge computing in response to the computing power requirements of artificial intelligence development, and it offers the industry edge computing products and solutions that meet a variety of AI requirements. UnixCloud launched the 10G network card business at the same time, offering four-port and two-port 10G optical fiber network cards based on the Mucse-based network controller N10.